Public reports
Each report is one run of a batch comparison, published by the person who ran it. Every number comes from that run.
Text in images: which model spells it right?
B (Nano Banana 2 (Gemini 3.1 Flash Image)) passed 67% of judged outputs, 33 points below A, at 5.3 times A's cost per call.
Three design briefs: which model invents facts?
C (GPT-5.6 Terra) passed 100% of judged outputs, 67 points above A, at 30% of A's cost per call.